Your web-browser is very outdated, and as such, this website may not display properly. Please consider upgrading to a modern, faster and more secure browser. Click here to do so.

Constrained Optimum

Research tumblr of a econ grad student - this is where I write about whatever I'm working on. Expect informal institutions, political violence, and data visualization.
May 30 '12
Posting this here so I can link to it in my next GPJ piece.

Posting this here so I can link to it in my next GPJ piece.

May 2 '12
Posting this here so I have somewhere to get the image for an upcoming Global Policy piece.

Posting this here so I have somewhere to get the image for an upcoming Global Policy piece.

Apr 18 '12
Here’s some data from the EPO and the USPTO on recent patent activity in renewable energy technology. I’m mostly just posting this here to include it in a Global Policy piece which will be up later this week.

Here’s some data from the EPO and the USPTO on recent patent activity in renewable energy technology. I’m mostly just posting this here to include it in a Global Policy piece which will be up later this week.

Mar 18 '12
Actually pretty pleased with this instrument I’m using for electrification. It tries to predict how much electricity a country produces using characteristics like river volume, population density, and number of land borders with other countries - and it does a pretty decent job. The correlation here is around 0.78.

Actually pretty pleased with this instrument I’m using for electrification. It tries to predict how much electricity a country produces using characteristics like river volume, population density, and number of land borders with other countries - and it does a pretty decent job. The correlation here is around 0.78.

Mar 5 '12
So here’s the data on skill-biased technological change in Costa Rica. Basically, here we’re looking at how the structure of the labour market changes to favour more highly-skilled people in terms of wages and employment. All the data here has been normalized to the lowest-education level of wage and employment in any given year, so all you need to look at is the relative magnitudes and the sign.
Not going to lie, I was not expecting the data here to be quite so noisy. It looks like the rate of wage growth really fluctuates heavily in industrializing countries, which I guess I should have expected. Not entirely sure what the ramifications are going to be for the study I’m trying to do here.

So here’s the data on skill-biased technological change in Costa Rica. Basically, here we’re looking at how the structure of the labour market changes to favour more highly-skilled people in terms of wages and employment. All the data here has been normalized to the lowest-education level of wage and employment in any given year, so all you need to look at is the relative magnitudes and the sign.

Not going to lie, I was not expecting the data here to be quite so noisy. It looks like the rate of wage growth really fluctuates heavily in industrializing countries, which I guess I should have expected. Not entirely sure what the ramifications are going to be for the study I’m trying to do here.

Mar 3 '12
Can’t help but feel like this GMM model has gotten away from me a little. I’m a little worried about degrees of freedom here; this is eight equations which will have forty or fifty parameters, and I only have a few hundred data points. Unless I get very lucky these are going to be some very unfortunate error bars.
Still, I can’t think of a good way to measure fuel efficiency (which is totally unobservable) without a pretty broad model like this. So this is going to have to do for now.

Can’t help but feel like this GMM model has gotten away from me a little. I’m a little worried about degrees of freedom here; this is eight equations which will have forty or fifty parameters, and I only have a few hundred data points. Unless I get very lucky these are going to be some very unfortunate error bars.

Still, I can’t think of a good way to measure fuel efficiency (which is totally unobservable) without a pretty broad model like this. So this is going to have to do for now.

Mar 1 '12
Trying to calculate the skill premium of wages, and realized that something strange was happening. Turns out my dataset (namely, the excellent and thorough SEDLAC) doesn’t correct for inflation and revaluation. This is going to be a problem, since for many of the countries in the data (like Brazil, shown above) currency fluctuations totally overpower any other trend.

Trying to calculate the skill premium of wages, and realized that something strange was happening. Turns out my dataset (namely, the excellent and thorough SEDLAC) doesn’t correct for inflation and revaluation. This is going to be a problem, since for many of the countries in the data (like Brazil, shown above) currency fluctuations totally overpower any other trend.

Jan 13 '12

creating an index for skill-biased technological change

An increasingly-widely accepted conclusion in economics (especially in the context of rapidly industrializing countries) is that in the last twenty or twenty-five years technological growth has been biased in favour of high-skilled workers - that is, the changes have made high-skilled workers more productive relative to low-skilled workers. Most people think this is primarily driven by greater computerization and automation, which replaces several (low-skill) workers doing relatively repetitive tasks with a single (high-skill) worker managing machinery that performs those tasks. (In this context, we’re basically defining your skill level as the number of years of education you’d need before you could start at the job. This is not really what “skill” means but in this context that’s what I’m talking about.)

This isn’t necessarily a bad thing, as it can raise everyone’s wages and overall standard of living (including low-skilled workers), but it does mean that wages are going to rise fastest for higher-skilled workers. Without policy intervention, this leads to sharply rising income inequality.

Anyway, I’m trying to write a paper on the role of certain kinds of infrastructure in skill-biased technological growth, but for the statistical end of things I need to come up with a good index for how much of the technological growth favours high-skilled workers over low-skilled, or vice versa. There’s obviously no one definite way to measure this, so I have to be sneaky and find something persuasive.

The standard way to deal with this is to use data from the United Nations Industrial Development Organization, and basically use the proportion of manufacturing wages which go to management, engineering staff, and other white-collar employees (as opposed to people who actually work in production) as a proxy for the premium on skilled labour. Unfortunately, the relevant data set ends in the early 1990s, and also I’m not entirely sure the white-collar/blue-collar divide accurately captures the premium on skilled labour. Like, I’m entirely willing to believe that a technician at a pharmaceutical plant might have had to spend way more time and money acquiring their skills than a manager at a meat-packing plant, you know? So, for this paper I’m going to be coming up with a new measure of skill-biased technological change.

It works something like this: I’m going to take existing measures of how R&D-intensive various manufacturing sectors are, and for each country in each year I’m going to regress the share of total manufacturing wages on the level of R&D intensiveness. If the level of R&D intensiveness shows a positive year-on-year shift, then that’s a sign that technological change is biased towards high-skilled workers. If it’s basically the same every year, then that’s a sign that technological change is affecting everyone more or less equally.

Does this sound like a reasonable measure? Or is it unconvincing to associate R&D intensiveness with the level of skill associated with manufacturing? (For what it’s worth, the ordering looks pretty reasonable, with high-precision manufacturing like aerospace at the top and relatively bulk goods like textiles near the bottom.) As always, feedback is appreciated.

Nov 21 '11
Here are the payoffs as a function of bids for a heterogeneous case. As you can see, something seriously suboptimal is happening here; it’s entirely unreasonable to expect anyone to maximize over payoff curves that look like this. I’m seriously hoping that this is just a mistake in my math, or else the whole concept of level-k equilibria doesn’t apply here.

Here are the payoffs as a function of bids for a heterogeneous case. As you can see, something seriously suboptimal is happening here; it’s entirely unreasonable to expect anyone to maximize over payoff curves that look like this. I’m seriously hoping that this is just a mistake in my math, or else the whole concept of level-k equilibria doesn’t apply here.

Tags: rosca
Nov 21 '11
Having a big problem that shouldn’t be a problem at all in theory but turns out to matter in practice.
These are payoff curves for a rosca with heterogeneous participants; the different curves are for different numbers of participants. The bottom curve is the case with just two participants. Note how it’s way less strongly peaked than any of the others? This is serious issue for experimental implementation; flat payoff curves mean that participants will be basically indifferent across any bid from 0.7 to 0.85, which makes it hard to figure out what they’re doing when I run the stats. Ideally, I want all the payoff curves to be as sharply peaked as the ones at the top.
Since I can control all the parameters of the experiment, I’m messing around trying to find something that makes all the payoff curves reliably steep. Unfortunately, the only good way I’ve found so far is to set their discount rate really high - like, discounting each successive period by a factor of 0.5. To implement this, there would have to be a 50/50 chance of stopping the experiment at each period (because that’s the only good way to implement this sort of future discounting). However this would make it basically impossible to get good data on a 10-member rosca, since the odds of actually getting to the end would be 1/1024. Obviously suboptimal.
Alternately I can argue it doesn’t really matter, if the odds of getting down to the last one or two participants without the experiment stopping first is really low. Even if I stop the experiment with only a 25% chance at each period, the odds of getting down to the tenth period are around 5%. Definitely not sure how to proceed here, though.

Having a big problem that shouldn’t be a problem at all in theory but turns out to matter in practice.

These are payoff curves for a rosca with heterogeneous participants; the different curves are for different numbers of participants. The bottom curve is the case with just two participants. Note how it’s way less strongly peaked than any of the others? This is serious issue for experimental implementation; flat payoff curves mean that participants will be basically indifferent across any bid from 0.7 to 0.85, which makes it hard to figure out what they’re doing when I run the stats. Ideally, I want all the payoff curves to be as sharply peaked as the ones at the top.

Since I can control all the parameters of the experiment, I’m messing around trying to find something that makes all the payoff curves reliably steep. Unfortunately, the only good way I’ve found so far is to set their discount rate really high - like, discounting each successive period by a factor of 0.5. To implement this, there would have to be a 50/50 chance of stopping the experiment at each period (because that’s the only good way to implement this sort of future discounting). However this would make it basically impossible to get good data on a 10-member rosca, since the odds of actually getting to the end would be 1/1024. Obviously suboptimal.

Alternately I can argue it doesn’t really matter, if the odds of getting down to the last one or two participants without the experiment stopping first is really low. Even if I stop the experiment with only a 25% chance at each period, the odds of getting down to the tenth period are around 5%. Definitely not sure how to proceed here, though.